Unibus disk controller with modern storage

[Eric Smith]

On Wed, Oct 19, 2016 at 4:48 PM, shadoooo <shadoooo at gmail.com> wrote:

> The board itself wouldn't be cheap at all, because PCB would be big,
>

True. From a Chinese vendor such as PCBway, a DEC quad size double-sided
PCB without ENIG (immersion) gold surface finish but without hard-gold edge
fingers costs $15.10 each in quantity 10, and a hex side PCB costs $20.40
each.  However, the ENIG gold is quite thin and won't withstand very many
backplane insertions.  The cheap PCB vendors don't offer hard gold edge
fingers. Jon Elson pointed out that E-TekNet in Arizona does offer hard
gold at better prices than some fab houses, but still a lot more than the
cheap Chinese PCBs.

I prefer NOT to use ENIG, as I find HASL tin-lead better for hand assembly,
though the lead is a problem due to RoHS regulations in much of the world
(but not in USA). I haven't tried HASL lead-free.


> and because FPGA aren't cheap either.
>

Xilinx XC6SLX9-2TQG144C is probably big enough for such things, and only
costs $16.52 in quantity 1 from Digi-Key.  It's in a TQFP, so not *too*
difficult to deal with. It has essentially 11,440 logic cells (5-LUT with
FF)*,  32 block RAMs of 18 Kbits each, and 102 3.3V I/O pins. It's not 5V
tolerant, but no modern parts are. 5V tolerance can be achieved in many
cases by the use of series resistors, but I like using the TI
SN74CBTD3861DGVR bus switch/voltage clamp, which can make 10 I/O pins 5V
tolerant at a cost of $0.62 in quantity 1. The bus switch is advantageous
over series resistors because it doesn't add much series resistance to pins
that are being used as outputs, and it has a maximum propagation delay of
0.25ns.

If you need more FPGA capability, the XC7A15T-1FTG256C has substantially
more resources, and costs $25.69 in quantity 1.  It's in a 256 ball BGA, so
somewhat harder to deal with, and needs at least a four layer board,
possibly even six layer.  However, it has 20,800 logic cells (5-LUT with
FF), 25 block RAMs of 36 Kbits each, and 170 3.3V I/O pins.

If you need more resources than that, it turns out that the XC7A15T,
XC7A35T, and XC7A50T are actually all the same die, just factor-programmed
with a different device ID. The 35T and 50T have double and triple the
logic cells and block RAMs of the 15T. The Vivado FPGA toolchain
artificially limits the resource usage of the two smaller parts, but
doesn't actually restrict which specific logic cells and block RAMs are
used, which means that the 15T and 35T have silicon that passes the factory
testing for ALL resources, not just 1/3 or 2/3 of them. It's been verified
that one can change the device ID in the bitstream, disable bitstream CRC
checking, and use the smaller part as a larger part. I don't like disabling
the CRC, because it serves to protect the FPGA from damage if the bitstream
has been corrupted, so I wrote a program that can both change the device ID
and recompute the CRC. I don't yet have a board with a 15T or 35T to test
with, but I'll release the program as open source once I do.

Because going to a 4 layer or 6 layer board is quite expensive when the
board size is large, e.g., DEC quad or hex size, I think it makes sense to
put the FPGA, its configuration flash, voltage regulators, and the bus
switches for 5V tolerance on a daughterboard. I've been working on such a
design, with two 96-pin DIN connectors for connection to the main board,
and at least 160 GPIOs available. The main board can then be just
double-sided, and possibly even 100% through-hole, for people that don't
want to hand-assemble boards with surface-mount components. The drawback is
that it will occupy more than one backplane slot due to the height.

There's still the problem that no current production bus interface ICs meet
Unibus, Qbus, or Omnibus specifications. Surplus DS8641 chips are
technically the best choice. With modern chips, it's necessary to use
separate drivers and receivers, and still difficult to meet the full specs.
Compromising the specs is possible but may make things unreliable in a
large system (multiple backplanes with cables, and many other cards).

Eric

* Xilinx has some other confusing definition of a "logic cell" for
marketing purposes, which is not the same as a LUT+FF. Oddly enough, their
marketing logic cell count is actually lower than a sensible accounting,
which is the opposite of how FPGAs used to be rated in exaggerated gate
counts, which we derided as "marketing gates".

The 6-series and 7-series actually have 6-LUTs with 2 FF each, but they can
be used as two 5-LUTs with 1 FF each, which is how I count them for
assessing the FPGA's logic capacity.